There is a dramatic and growing shortage of highly skilled welders, accentuated by the fact that manufacturing complexity and production volumes are rising. As a result, global use of robotic welding is expanding rapidly. However, current welding robots are not as adaptive and creative as human welders in performing complex welding tasks that require sophisticated skills. This award supports fundamental research on advancing the robotic capabilities needed to realize fully robotic automation of complex welding tasks. The research will endow collaborative welding robots with sophisticated welding knowledge, expert intelligence, and an interactive learning capability to enable them to address dynamic welding scenarios. The research results will both enhance the scientific base for robotic control and facilitate the realization of fully automatic, robotic, and intelligent manufacturing. The research involves several disciplines, including welding, process monitoring, data visualization, machine learning, optimization, and robotic control. That multi-disciplinary approach will broaden the participation of students from diverse backgrounds in research, and the knowledge gained will be incorporated in curricula in robotic and intelligent manufacturing.
The double-electrode, gas metal arc welding process is complex, requiring intense collaboration between expert welders. As a result, the robotic automation of such a complex welding process requires advances in the scientific base of robotic perception, learning, and control. The project will research advanced methods for the extraction of expert welding-domain knowledge and the quantification and interpretation of that knowledge for use by collaborative robots, thereby equipping collaborative welding robots to perform complex welding tasks. To realize that goal, the research team will: 1) build an immersive virtual reality system with a three-dimensional rendering of the weld pool and arc that can characterize the weld scene and record human operations, 2) use an explainable recurrent convolutional neural network to perform causal analysis of the torch manipulation of human welders to obtain its relationship to dynamic weld pool/arc evolution, 3) generalize the results in terms of human heterogeneity by using transfer learning to extract common latent knowledge from different human welders, and 4) develop an interactive learning module that allows collaborative robots to be supervised by on-site human welders through the reinforcement learning-based perception of language instructions and human gestures.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
